Google spots an AI agent turning security patches into exploits with no human involved
Google's Threat Intelligence Group says it found exploit_agent, a multi-agent framework that diffs patched web software, writes exploits, and runs them on its own, the first fully autonomous pipeline it has seen in the wild.
2 min readJubaer Jami
Google Threat Intelligence Group (GTIG) published its latest quarterly AI threat tracker on November 16, 2026, together with Mandiant. Its headline finding: for the first time, the group says it has seen attackers use LLM frameworks to find and exploit vulnerabilities from start to finish without a person in the loop.
What Google found
The evidence came from an exposed staging server. On it was a multi-agent framework the operator called exploit_agent, wired to several frontier and locally hosted language models. According to Google, the system watches popular web application software, compares code before and after patches are announced, works out what weakness a fix likely closes, drafts an exploit, and then runs it automatically.
The same server held data taken from victims and scripts for post-compromise work: internal reconnaissance, brute-forcing, installing webshells, and deleting webshells left by other attackers. Victim telemetry pointed to an interest in French media, entertainment, and education websites. One example in the report shows the framework examining access-control code in the Spectra WordPress plugin. Configuration comments were written in Russian.
SiliconANGLE notes that in the previous edition of the tracker, GTIG said it had not yet seen fully autonomous attack pipelines used against real targets.
Why it matters
Google's argument is that patched, publicly known bugs are an efficient target for AI: the source material is finite and the patch itself points to the flaw. That shrinks the time defenders have between a fix being published and an exploit appearing. The report also describes attackers adopting open-source agentic pentesting frameworks such as CAIRN, CyberStrikeAI, and PentAGI, and selling stolen access to AI models.
The findings rest on Google's own investigation; the report does not say how often the generated exploits succeeded.
Sources
This post relies on the documents below. Open them to check the reporting behind it.
- 01
- 02